In doing design of a physical system it is often useful to reduce the computational cost of evaluating design alternatives by using a range of simluation/analysis models with different degrees of accuracy and computational cost. In the early design stage, we would trade off the accuracy of a model in exchange for its efficiency. In the final design where high degree of accuracy is required, then we switch to an accurate but costly model. We are studying the problem of automatically constructing such models in the domain of heat transfer. This domain is different from most of those studied by AI/Modelling research because both lumped and distributed continuous behavior are involved. The models in this domain ranges from simple algebraic equations to complex partial differential equations. Our focus is on models to be used by the evaluation step of an iterative parameter design process.